search

dvlab-research / outpainting_srn

Wide-Context Semantic Image Extrapolation, CVPR2019
MIT License
132 stars 32 forks source link

train problem #5

Open zhangbaijin opened 5 years ago

zhangbaijin commented 5 years ago

for example,if i use celebahq,i should follow the cmd：python train.py --dataset celebahq --data_file /data/yiwang/download/celeba_hd_256/train.txt --gpu_ids 0 --max_iters 80000 --g_cnum 64 --pretrain_network 1 --batch_size 8 --model srn --feat_expansion_op subpixel --random_mask 1 --random_crop 0 --use_cn 1

what's the train.txt? how should i get it when the dataset is Paris-Streetview?

shepnerd commented 5 years ago

The mentioned train.txt is a file storing the full paths of the training images, like: /data/face/0x1.png /data/face/0x2.png ... /data/face/0f4.png

To generate the corresponding training file of a given dataset, you can traverse all its training data and write their full paths into a file.

zhangbaijin commented 5 years ago

ok！thank you ,and another problem occured, No such file or directory: './checkpoints/20190717-153620_./dataset/Paris_Streetview/paris_train_original_srn_subpixel_contextual_b8_s936x537_gc32_dc64_rand-mask'

shepnerd commented 5 years ago

It seems some input options are invalid. The input for the option --dataset is just a dataset name like 'paris_streetview' instead of a path. If given a path, the creation of checkpoint folder would fail.

zhangbaijin commented 5 years ago

thank you for your answer!but the problem still happen,my paris-streetview include 3 folders and 1 file,paris_eval_corrupted, paris_eval_gt, paris_train_original and paris-streetview.txt . The paris_train_original image shape is 936X537. So i follow the cmd:

python train.py --dataset paris_streetview --data_file ../paris_streetview.txt --gpu_ids 0 --img_shapes 936,537 --pretrain_network 0 --batch_size 8 by the way,my paris-streetview like this: /dataset/paris-streetview/48.842502_2.344968270-004.JPG /dataset/paris-streetview/48.842502_2.34496890-004.JPG /dataset/paris-streetview/48.842513_2.343543270-004.JPG /dataset/paris-streetview/48.842513_2.34354390-004.JPG /dataset/paris-streetview/48.842529_2.369006270-004.JPG /dataset/paris-streetview/48.842529_2.36900690-004.JPG

shepnerd commented 5 years ago

About the experimental settings, you can refer to our paper. We use randomly cropped 256x512 (hxw) images from paris-streetview for training and 256x256 images in eval_gt for evaluation. And we set --random_mask to 0 for the view expansion.